SODA: Robust Training of Test-Time Data Adaptors

TL;DR

SODA introduces a robust test-time data adaptation method using zeroth-order optimization that leverages high-confidence pseudo-labels to improve model performance under distribution shifts without accessing model parameters.

Contribution

The paper proposes SODA, a novel pseudo-label-robust data adaptation technique that enhances zeroth-order optimization for test-time adaptation under privacy constraints.

Findings

SODA significantly improves model accuracy under distribution shifts.

SODA outperforms existing test-time adaptation methods.

SODA does not require access to model parameters.

Abstract

Adapting models deployed to test distributions can mitigate the performance degradation caused by distribution shifts. However, privacy concerns may render model parameters inaccessible. One promising approach involves utilizing zeroth-order optimization (ZOO) to train a data adaptor to adapt the test data to fit the deployed models. Nevertheless, the data adaptor trained with ZOO typically brings restricted improvements due to the potential corruption of data features caused by the data adaptor. To address this issue, we revisit ZOO in the context of test-time data adaptation. We find that the issue directly stems from the unreliable estimation of the gradients used to optimize the data adaptor, which is inherently due to the unreliable nature of the pseudo-labels assigned to the test data. Based on this observation, we propose pseudo-label-robust data adaptation (SODA) to improve the performance of data adaptation. Specifically, SODA leverages high-confidence predicted labels as reliable labels to optimize the data adaptor with ZOO for label prediction. For data with low-confidence predictions, SODA encourages the adaptor to preserve data information to mitigate data corruption. Empirical results indicate that SODA can significantly enhance the performance of deployed models in the presence of distribution shifts without requiring access to model parameters.